1. Segmentation with non-linear constraints on appearance, complexity, and geometry Yuri Boykov Western Univesity Andrew Delong Olga Veksler Lena Gorelick Frank Schmidt Anton Osokin Hossam Isack IPAM February 2013

2. Overview Standard linear constraints on segments – intensity log-likelihoods, volumetric ballooning, etc. 1. Basic non-linear regional constraints – enforcing intensity distribution (KL, Bhattacharia, L 2 ) – constraints on volume and shape 2. Complexity constraints (label costs) – unsupervised and supervised image segmentation, compression – geometric model fitting 3. Geometric constraints – unsupervised and supervised image segmentation, compression – geometric model fitting (lines, circles, planes, homographies, motion,…)

3. Image segmentation Basics 3 S

4. Linear appearance of region S Examples of potential functions f Log-likelihoods Chan-Vese Ballooning

5. Part 1 Basic non-linear regional functionals

6. Standard Segmentation Energy 6 Fg Bg Intensity Probability Distribution Target Appearance Resulting Appearance

7. Minimize Distance to Target Appearance Model 7 Non-linear harder to optimize regional term Non-linear harder to optimize regional term

8. – appearance models – shape non-linear regional term 8

9. Related Work Can be optimized with gradient descent – first order approximation 9 Ben Ayed et al. Image Processing 2008, Foulonneau et al., PAMI 2006 Foulonneau et al., IJCV 2009 Ben Ayed et al. Image Processing 2008, Foulonneau et al., PAMI 2006 Foulonneau et al., IJCV 2009 We use higher-order approximation based on trust region approach

10. a general class of non-linear regional functionals

11. Regional Functional Examples Volume Constraint 11

12. Bin Count Constraint 12 Regional Functional Examples

13. Histogram Constraint 13

14. Regional Functional Examples Histogram Constraint 14

15. Regional Functional Examples Histogram Constraint 15

16. Shape Prior Volume Constraint is a very crude shape prior Can be generalized to constraints for a set of shape moments m pq 16

17. Volume Constraint is a very crude shape prior Shape Prior 17

18. Shape Prior using Shape Moments m pq 18

19. Shape Prior Constraint Shape Prior using Shape moments 19 Dist(),

20. Optimization of Energies with Higher-order Regional Functionals 20

21. Gradient Descent (e.g. level sets) Gradient Descent First Order Taylor Approximation for R(S) First Order approximation for B(S) (“curvature flow”) Robust with tiny steps – Slow – Sensitive to initialization 21 http://en.wikipedia.org/wiki/File:Level_set_method.jpg Ben Ayed et al. CVPR 2010, Freedman et al. tPAMI 2004 Ben Ayed et al. CVPR 2010, Freedman et al. tPAMI 2004

22. Energy Specific vs. General Speedup via energy- specific methods – Bhattacharyya Distance – Volume Constraint We propose – trust region optimization algorithm for general high-order energies – higher-order (non-linear) approximation Ben Ayed et al. CVPR 2010, Werner, CVPR2008 Woodford, ICCV2009 Ben Ayed et al. CVPR 2010, Werner, CVPR2008 Woodford, ICCV2009 22

23. General Trust Region Approach An overview The goal is to optimize Trust region Trust Region Sub-Problem First Order Taylor for R(S) Keep quadratic B(S) First Order Taylor for R(S) Keep quadratic B(S) 23

24. General Trust Region Approach An overview The goal is to optimize Trust Region Sub-Problem 24

25. Solving Trust Region Sub-Problem Constrained optimization minimize Unconstrained Lagrangian Formulation minimize Can be optimized globally u using graph-cut or convex continuous formulation L 2 distance can be approximated with unary terms [Boykov, Kolmogorov, Cremers, Delong, ECCV’06] L 2 distance can be approximated with unary terms [Boykov, Kolmogorov, Cremers, Delong, ECCV’06] 25

26. Approximating distance [BKCD – ECCV 2006]

27. Trust Region Standard (adaptive) Trust Region – Control of step size d Lagrangian Formulation – Control of the Lagrange multiplier λ 27 λ

28. Spectrum of Solutions for different λ or d 28 Newton step “Gradient Descent” Exact Line Search (ECCV12) Newton step “Gradient Descent” Exact Line Search (ECCV12)

29. Volume Constraint for Vertebrae segmentation Log-Lik. + length + volume Fast Trust Region Initializations Log-Lik. + length 29

30. Appearance model with KL Divergence Constraint Init Fast Trust Region “Gradient Descent” Exact Line Search 30 Appearance model is obtained from the ground truth

31. Appearance Model with Bhattacharyya Distance Constraint 31 “ “ Fast Trust Region “Gradient Descent” Exact Line Search Appearance model is obtained from the ground truth

32. Shape prior with Tchebyshev moments for spine segmentation 32 Log-Lik. + length + Shape Prior Fast Trust Region Second order Tchebyshev moments computed for the user scribble

33. Part 2 Complexity constraints on appearance

34. Segment appearance ? sum of log-likelihoods (linear appearance) – histograms – mixture models assumes i.i.d. pixels with constraints – based on information theory (MDL complexity) – based on geometry (anatomy, scene layout) now: allow sub-regions (n-labels)

35. Natural Images: GMM or MRF? 35 are pixels in this image i.i.d.?NO!

36. Natural Images: GMM or MRF? 36

37. Natural Images: GMM or MRF? 37

38. Natural Images: GMM or MRF? 38

39. Binary graph cuts 39 [Boykov & Jolly, ICCV 2001] [Rother, Kolmogorov, Blake, SIGGRAPH 2004]

40. 40 [Boykov & Jolly, ICCV 2001] [Rother, Kolmogorov, Blake, SIGGRAPH 2004] Binary graph cuts

41. 41 [Boykov & Jolly, ICCV 2001] [Rother, Kolmogorov, Blake, SIGGRAPH 2004] Binary graph cuts

42. Objects within image can be as complex as image itself Where do we draw the line? A Spectrum of Complexity 42 MRF?GMM? Gaussian? object recognition??

43. Single Model Per Class Label 43 Pixels are identically distributed inside each segment

44. Multiple Models Per Class Label 44 Now pixels are not identically distributed inside each segment

45. Multiple Models Per Class Label 45 Now pixels are not identically distributed inside each segment Our Energy ≈ Supervised Zhu & Yuille!

46. 46 boundary length MDL regularizer + color similarity +  -expansion (graph cuts) can handle such energies with some optimality guarantees [IJCV’12,CVPR’10] Leclerc, PAMI’92 Zhu & Yuille. PAMI’96; Tu & Zhu. PAMI’02 Unsupervised clustering of pixels

47. (unsupervised image segmentation) Fitting color models information theory (MDL) interpretation: = number of bits to compress image I losslessly each label L represents some distribution Pr(I|L)

48. (unsupervised image segmentation) Fitting color models Label costs only Delong, Osokin, Isack, Boykov, IJCV 12 (UFL-approach)

49. (unsupervised image segmentation) Fitting color models Spatial smoothness only [Zabih & Kolmogorov, CVPR 04]

50. (unsupervised image segmentation) Fitting color models Spatial smoothness + label costs Zhu & Yuille, PAMI 1996 (gradient descent) Delong, Osokin, Isack, Boykov, IJCV 12 (a-expansion)

51. Fitting planes (homographies) Delong, Osokin, Isack, Boykov, IJCV 12 (a-expansion)

52. Fitting planes (homographies) Delong, Osokin, Isack, Boykov, IJCV 12 (a-expansion)

53. Back to interactive segmentation 53

54. 54 segmentationcolour models“sub-labeling” EMMCVPR 2011

55. Main Idea Standard MRF: Two-level MRF: 55 object MRF GMMs background MRF image-level MRF object GMMbackground GMM image-level MRF unknown number of labels in each group!

56. Our multi-label energy functional 56 Penalizes number of GMMs (labels) – prefer fewer, simpler models – MDL / information criterion regularize “unsupervised” aspect Discontinuity cost c is higher between labels of different categories data costssmooth costs label costs GMMs set of labels L background labels object labels two categories of labels (respecting hard-constraints)

57. More Examples 57 standard 1-level MRF 2-level MRF

58. More Examples 58 2-level MRF standard 1-level MRF

59. Beyond GMMs 59 GMMs GMMs + planes plane GMMs only

60. Part 3 Geometric constraints on sub-segments

61. Sub-labels maybe known apriori -known parts of organs or cells -interactivity becomes optional Geometric constraints should be added - human anatomy (medical imaging) - or known scene layout (computer vision) Towards biomedical image segmentation…

62. Illustrative Example We want to distinguish between these objects 62 input what mixed model giveswhat we want mixed appearance model:

63. main ideas – sub-labels with distinct appearance models (as earlier) – basic geometric constraints between parts -inclusion/exclusion -expected distances and margins For some constraints globally optimal segmentation can be computed in polynomial time 63

64. Illustrative Example Model as two parts: “light X contains dark Y” 64 Center object the only object that fits two-part model (no localization)

65. ‘layer cake’ a-la Ishikawa’03 Our Energy 65 variables Let over objects L and pixels P

66. Surface Regularization Terms Standard regularization of each independent surface 66 Let over objects L and pixels P

67. Geometric Interaction Terms Inter-surface interaction 67 Let over objects L and pixels P

68. So What Can We Do? Nestedness/inclusion of sub-segments ICCV 2009 (submodular, exact solution) Spring-like repulsion of surfaces, minimum distance ICCV 2009 (submodular, exact solution) Spring-like attraction of surfaces, Hausdorf distance ECCV 2012 (approximation) Extends Li, Wu, Chen & Sonka [PAMI’06] – no pre-computed medial axes – no topology constraints 68

69. Applications Medical Segmentation – Lots of complex shapes with priors between boundaries – Better domain-specific models Scene Layout Estimation – Basically just regularize Hoiem-style data terms [4] 69

70. Application: Medical 70 our result input

71. Application: Medical 71 full body MRI two-part model

72. Application: Medical 72 full body MRI

73. Conclusions linear functionals are very limited – be careful with i.i.d. (histograms or mixture models) higher-order regional functionals use sub-segments regularized by – complexity or sparcity (MDL, information theory) – geometry (based on anatomy) literature: ICCV’09, EMMCVPR’11, ICCV’11,, IJCV12, ECCV’12